Research On The Prediction Method Of The Ridge Index Of The Western Pacific Subtropical High Based On Shallow Neural Network

The ridge line of the western Pacific subtropical high(WPSHRL),which is one of the general atmospheric circulation indexes,plays an important role in determining the shift of the summer rain belt in eastern China.At present,the forecast skill of the traditional statistical methods and numerical model methods predicting WPSHRL index remains unsatisfactory(Cor< 0.5)at lead times of more than one season.In recent years,some studies have shown that using machine learning methods can significantly improve the skills of climate forecasting.This paper studies how to use machine learning methods to predict the WPSHRL index and improve the forecasting skills of predicting the WPSHRL index in June at lead times of more than one season.Different from other "big data" fields,machine learning methods for climate prediction face two major problems: 1)Stability.The main reason is that the observation sample size is too small;2)Interpretability.The relationship between input predictors and output predictands presented by the prediction model cannot violate the existing climate background knowledge.This paper studies the machine learning method of forecasting the June WPSHRL index based on the latest autumn and winter sea surface temperature(SST).Considering the small observed sample size(only a few decades),a simple shallow neural network model was selected to extract the non-linear relationship between input predictors(SST)and target predictands(WPSHRL).In addition,through the SST composites in the high and low WPSHRL index years,some training samples are manually extended to suppress the over-fitting problem in climate small sample learning.In order to avoid the sample imbalance in data set division,the "leave-one-out" method and "early stopping" threshold are used to preliminarily train the model,and the optimal hyperparameters are selected.According to certain rules,the weight initialization seeds are screened so as to improve the model stability and prediction skills.The evaluation of the forecast model shows that the linear correlation coefficient between the predictions of the WPSHRL index and their corresponding observations is greater than 0.6,and about three-fifths of the observed abnormal years(the years with an obviously high or low WPSHRL index)are successfully predicted.The interpretability analysis shows that the spatial characteristics of the sensitivity of the output predictands(WPSHRL)to the input predictors(SST)are consistent with the climate background knowledge(i.e.the linear correlation between the ridge and the previous SST),and the relationship between the input and output of the forecast model is interpretable.The stability analysis shows that the forecast skill of the model is very stable(almost repeatable),and the predictions are also relatively stable.In conclusion,the forecast model has certain practical application value and can be used for climate prediction services.